EmmaLeonhart/reservoir-agent-gpt2-batch

Reservoir Agent batch — gpt2

A batch of 4 reservoir agents (different fixed-random reservoir seeds) trained on the cross-pass recall task. A reservoir agent is a new model type: a pretrained transformer with a fixed reservoir brain-surgeried in (attended, cross-pass-stateful, RNN-like) — see the project and RESERVOIR_AGENTS.md.

The whole population is published, not just the winner. Reservoir performance is stochastic in the seed; the suboptimal models are kept as signal for learning which reservoir properties survive selection. The recommended model is seed_1.

Population

rank	seed	recall	loss_end	pr_frac	recommended
0	seed_1	1.00	0.000	0.114	yes
1	seed_0	1.00	0.001	0.114
2	seed_2	1.00	0.001	0.113
3	seed_3	1.00	0.200	0.112

Use

Each seed_<n>/ is a complete loadable reservoir agent. Load the recommended one:

from huggingface_hub import snapshot_download
from reservoir.persist import load_reservoir_model
path = snapshot_download("EmmaLeonhart/reservoir-agent-gpt2-batch")
lm = load_reservoir_model(f"{path}/seed_1")

batch_manifest.json records the ranking + each seed's score and reservoir-dynamics signal.